Abstract: Nuclear power valve has the characteristics of high reliability and long lifetime, and its failure data have obvious small sample problem. The nuclear power valve failure are caused by loss factors of impact, shaking, abrading and corrosion. The failure probability has time trend, and the p will raise（decrease） with the increasing of time. The Jeffreys prior model of invariable p can not reflect the time trend of p very well, so it can not analyze the time-varying characteristics of p. This paper built the Generalized Linear Model（GLM） for valves failure data having Binomial distribution, analyzed the time trend of p inspected the model through posterior predictive distribution, and assessed the model ability of replicating observed data by graph inspection and Bayesian chi-square. The results showed that GLM had well fit index and was more propitious to assess the failure probability p of valve.